vjp

Julia

vjp!(ḡ::AbstractVector, jac::JacobianEvaluator, λ::AbstractVector, u::AbstractVector, p, t::Real)

vjp!(ḡ, jac::JacobianEvaluator, λ, u, p, t) -> ḡ

The transposed product ḡ += Jᵀ λ at (u, p, t), as a scatter-add over the band points: ḡ[col] += coef · λ[row] for every point, with no matrix (and no transpose) formed. λ is indexed like the rows (jac.rownames, the state vector) and ḡ like the columns (jac.colnames); ḡ ACCUMULATES — zero it first for a single product.

Source

Python

vjp(g, obj, lam: dict, model_name: str | None = None, wrt: str | None = None, states: dict | None = None, params: dict | None = None, t: float = 0.0, forcing=None)

The transposed product g += Jᵀ λ at a point, as a scatter-add over the band points: g[col] += coef · λ[row], with no matrix formed.

obj is a JacobianEvaluator (built once with prepare_jacobian ; its axis and forcing set apply) or anything assemble_jacobian takes, prepared for this one call with wrt (default "forcing"). lam maps row (state cell) names to values; a row a band point needs that lam does not bind raises ValueError. g is a flat mutable sequence (list or numpy vector) over the canonical column ordinal of the "parameters" / "forcing" axis (jacobian_columns , cells column-major) or of "time" ([t]). g ACCUMULATES — zero it first for a single product. Returns g.

Source

Rust

pub fn vjp(file: &EsmFile, model_name: Option<&str>, wrt: Wrt, forcing: Option<&[String]>, env: &Env, lambda: &HashMap<String, f64>, gbar: &mut [f64]) -> Result<(), BandError>

The transposed product ḡ += Jᵀ λ at a point, as a scatter-add over the band points: ḡ[col] += coef · λ[row] for every point, with no matrix (and no transpose) formed. lambda is keyed by row cell name ("c[2]", "S"); a row it does not bind is an error. gbar is indexed by the canonical column ordinal of a parameters or forcing axis (crate::jacobian::column_names of crate::jacobian_columns ) and ACCUMULATES — zero it first for a single product. This prepares the evaluator for one call; a loop calls JacobianEvaluator::vjp on one prepare_jacobian instead.

Source · Rust docs